Generalization of IMPACT to Unseen Instruction Wording

Determine whether IMPACT generalizes its language-based task selection to instruction wording unseen during training and evaluation.

Background

IMPACT is evaluated by changing instructions while keeping LIBERO-Goal environments fixed, demonstrating selection among familiar goals. However, the evaluation does not test whether the policy can interpret novel linguistic formulations of those goals. Generalization to unseen wording therefore remains an unresolved empirical question.

References

This intervention supports language-based selection among familiar goals; generalization to unseen wording and the individual contributions of the two conditioning paths remain untested.

— vla.simd: Efficient CPU Inference for Language-Conditioned Manipulation  (2609.24274 - Nguyen et al., 21 Sep 2026) in Section 5, paragraph "Instruction following"